Introduction - L3 Production
The present chapter examines the impact of two previously known languages on the production of a novel language. In particular, it examines how Spanish-English bilinguals in both orders of acquisition produce German and French vowel and stop consonants. Models of L3 acquisition vary in their prediction regarding how previously known languages influence a third and suggest that typology, language status (L1 or L2) or language activation predict how either language will play a role in L3 development. All of these views have received empirical support to some degree, and there is still debate as to why these conflicting findings exist.
The typological view, for example, suggests that overall crosslinguistic similariy will result in whole-language influence. For example, a Spanish-English bilingual who learns Portuguese would be predicted to be influenced by their Spanish, given the close historical relationship between Spanish and Portuguese. While this view has been supported in syntax, it has not seen much empirical testing or support in phonology. In reality, an increasing number of studies are finding that both languages impact third language phonological development. However, it is still not well understood what impacts the relative degree of influence of the L1 or L2 and whether, while they may not cause whole language influence, whether the factors of language status and typology may still be at play.
In order to test this, the present study used a mirror-image design of participants and gave them words from two “third” languages produce (French and German). The mirror image design refers to two groups of speakers who speak the same two languages but acquired them in opposite order (e.g., L1 Spanish-L2 English and L1 English-L2 Spanish), The mirror image design is intended to reveal whether language status (order of acquisition) affects performance on a particular task, since these speakers theoretically have the same linguistic resources. If the groups perform differently, it could provide evidence that order of acquisition impacts access to linguistic representations in new language learning. The use of two “third” languages at first exposure was chosen to examine whether the same individual would be impacted by the same source language no matter what language they learn, or whether both the source languages known and the language being learned are important. Typological views of L3 acquisition would suggest that, given the historical relationship between French and Spanish and English and German respectively, that these languages would be produced more similarly. Language status views would predict that influence should not vary as a function of which L3 is being learned. Finally, views which suggest both languages impact L3 learning have not yet developed enough explanatory power to derive clear prediction related to Spanish-English bilinguals learning French and German.
L3 Production Studies
The findings in empirical studies of L3 phonological cross-linguistic influence have varied. One of the first studies to examine progressive influence of the L1 or L2 on L3 production was the seminal case study of @williams_language_1998. This study elicited the production of an adult L1 British English, L2 German, and L3 Swedish speaker in the L2 and L3 upon her arrival to Sweden. The speech samples were rated for native-likeness by native speakers of German and Swedish respectively, with low ratings (i.e. non-nativeness) being elaborated upon. In the event of non-native speech, the raters guessed where the speaker in the recording might be from. The informant was rated as having near native productions in her L2 German, while her L3 Swedish was rated as being non-native like and to be German-accented. The experiment was repeated after 6 months in Sweden, however, and the Swedish raters then judged the informant’s Swedish to be British English accented.
This study constituted evidence of an L2 status effect in the initial stages of L3 phonological acquisition, in which the second learned language influence L3 production, but also provided evidence that this effect diminishes as L3 proficiency increases. This notion has been called the ‘foreign language effect’ [@meisel_transfer_1983], which refers to the idea that speakers who learn a second non-native language are biased to sound unlike a native speaker of their native language.
The default L2 status effect has received some empirical support in the L3 phonology literature. In a study of global accent production, heavier L2 influence in L3 productions was found by L1 Polish, L2 German and L3 English speakers based on ratings of EFL instructors [@wrembel_l2-accented_2010]. Similar findings have also been reported in vowel production [@kamiyama_acquisition_2007] and vowel reduction and speech rhythm [@gut_cross-linguistic_2010]. L2 influence has also been found in VOT productions. @llama_influence_2010 examined L3 Spanish VOT production by French-English mirror-image bilingual groups and found that both groups had L2-like productions of the L3.
Other findings in L3 production, however, have yielded mixed results. Several studies have found that acoustic properties of the participants’ productions fall between L1 and L3 values, suggesting that both the L1 and the L2 have some influence on L3 productions, rather than solely one language. For instance, [@wrembel_vot_2014] measured VOT and aspiration in all languages of participants with two different language combinations: L1 Polish, L2 English, and L3 French; (2) L1 Polish, L2 English, and L3 German. The results showed that each language had a specific stop-value, and that the L3 VOT productions were intermediate, falling between the L1 and L2 values. Similarly, [@wrembel_cross-linguistic_2011] examined thirty-two learners of L3 French with L1 Polish and L2 English who were recorded reading lists of words in carrier phrases. As in previous studies [@wrembel_vot_2014], combined transfer from the L1 and the L2 in VOT productions was found. Findings of combined L1 and L2 influence in VOT productions were also reported by @wunder_phonological_2010 in L3 Spanish speakers, and by @blank_transferencia_2009 in L3 English speakers who spoke L1 Brazilian Portuguese and L2 French. Other studies have found an L1 influence on production despite L3 proficiency [@cabrelli_amaro_foreign_2012], or in advanced L3 learners [@llama_revisiting_2018]. Importantly, these studies report L3 VOT values which fall between L1 and L2 values. Following the suspicion that intermediate values might have to do with either sampling issues or proficiency effects, @parrish_relative_2021 examined Mexican Spanish-English bilinguals who produced voiceless stop-initial French words in isolation at first exposure to the language. The results found that the relative VOT of the L3 fell between their own L1 and L2 values, in line with previous research, and that suggested that intermediate values were less likely to have been seen in previous studies as a result of small samples or proficiency effects. However, a subsequent analysis of the data suggested that wide individual variation existed, in which some participants produced L3 French as L1 Spanish like, and other produced intermediate, L2-like values. This result suggests that higher samples could reveal group trends and provide better insights into individual variation in crosslinguistic influence, as opposed to assuming that a single group trend exists.
The present study
In order to examine the impact of order of acquisiton and language similarity on cross linguistic influence, the present study utilized word reading and word repetition tasks in order to elicit productions of vowels and stop consonants in 4 total languages in each participant (or 3 in the case of monolingual comparisons).
In particular, the present chapter is guided by the following research questions:
RQ1: When Spanish-English bilinguals produce French and German words at first exposure, to what degree will they be influenced by their known languages?
RQ2: Will source language impact vary as a function of segment (vowel or stop)?
The first research question was examined by extracting formant values from each of the vowels and measuring the VOT of each consonant in each language. These acoustic measures were recorded and are assumed to provide insight into a particular segment’s representation. That is, by comparing the acoustics of each sound, the present study extrapolates acoustics to phonological representation in each of the languages involved.
Vowel Production
Participants
A total of X participants participated in the present experiments and made up a total of 4 groups: L1 English bilingual (n = 20), L1 Spanish bilingual (n = 66), L1 English monolingual (n = 58) and L1 Spanish monolingual (n = 28). Just as in the perception experiments, the all participants completed the Bilingual Language Profile [@birdsong_bilingual_2012]. Figure @ref(fig:aoaprod) shows the age of Onset (A) and age of acquisition (B) by the L1 English and L1 Spanish bilingual groups. Figure @ref(fig:profprod) shows the self-rated proficiency of the same groups. All participants who answered ‘no’ to the question “Do you speak a language other than English and Spanish” were permitted to continue the experiment.
Self-rated spoken and eceptive proficiency by the L1 English and L1 Spanish bilingual groups
Materials
Word Reading Task
/pVf/, /fVf/
English: a, i, o, wedge faff, fiff, foff, fuff
Spanish: a, i, o, u faf, fif, fof, fuf paf, pif, pof, puf
Word Repitition Task
/pVf/, /fVf/
French: a, i, o, y faf, fif, fof, fuf paf, pif, pof, puf
German: a, i, o, y faf, fif, fof, fuf paf, pif, pof, puf
Procedure
Statistical Analysis
A series of Bayesian multilevel regression models were run to
determine whether there were differences in vowel height (F1) an and
frontness (F2) between groups. In each model, the formant was analyzed
as a function of language (4 levels for bilingual groups, 3 levels for
monolingual groups), and phoneme (4 levels). Random intercepts for token
and participant were also included to take into account the nested
structure of the data. Model priors were the default in
brms, a student’s T distribution with 3 degree of freedom.
All models were fit with 4000 iterations (1000 warm-up). Hamiltonian
Monte-Carlo sampling was carried out with 6 chains distributed between 8
processing cores.
Results
Forced Alignment
A total of 15019 data points were force aligned in the initial analysis using Webmaus basic (cite). In order to estimate the accuracy of the forced-alignment, about 3 percent of the bilingual data was randomly subset and hand-corrected (380 tokens). Subsetting was done using an R script. Figure @ref(fig:langcorr) shows the quantity of each phoneme in each language that was hand-corrected in the random subset.
Figure @ref(fig:handcorrect) shows the formant values of hand-corrected tokens and force-aligned tokens. An inspection of the figure suggests that /a/, wedge and /o/ were mostly accurately aligned. On the other hand, the tokens for /i/ and /u/ benefited from hand-correction.
## `summarise()` has grouped output by 'phoneme'. You can override using the
## `.groups` argument.
## `summarise()` has grouped output by 'phoneme'. You can override using the
## `.groups` argument.## Warning: position_dodge requires non-overlapping x intervals
Quantity of Hand-correced tokens per phoneme per language
Hand Corrected versus automated formant values of a subset of the data
Figure @ref(fig:rmdplot) shows the root mean square deviation between the hand-corrected values and force-aligned values. Root mean square deviation is an measurement of the difference between predicted and actual values. In the present study, these numbers represent Hertz for F1 and F2. Each numerical label in the figure represents the root mean square deviation, while the colors represent phonemes. The phonemes /i/ and /u/ have the largest root mean square deviation, showing, again, that they were the least accurately force-aligned.
## Warning: Ignoring unknown parameters: check_overlap## Warning: position_dodge requires non-overlapping x intervals
## position_dodge requires non-overlapping x intervals
Root Mean Square Deviation for F1 and F2
A total of 2005 tokens were removed from the dataset. A token was considered an outlier when it was more than plus or minus two standard deviations of the mean of the hand-corrected subset. Table @ref(tab:handcortab) in the appendix shows the mean and standard deviations of each phoneme in each language used to identify outliers in the main dataset. Figure @ref(fig:rmdremoved) shows the number of removed tokens compared to the root mean square deviation. The most removed phoneme was German /o/, in which 371 tokens were removed as outliers. Table @ref(tab:remtab) summarizes the total number of each phoneme removed. Figure @ref(fig:totaltokens) shows the the final number of tokens analyzed after removal of outliers (n = 13014) broken down by language and phoneme.
| phoneme | n |
|---|---|
| a | 54 |
| i | 496 |
| o | 676 |
| schwa | 278 |
| u | 501 |
## Warning: position_dodge requires non-overlapping x intervals
## position_dodge requires non-overlapping x intervals
Root Mean Square Deviation for F1 and F2 versus total removed tokens
## `summarise()` has grouped output by 'phoneme'. You can override using the
## `.groups` argument.
## `summarise()` has grouped output by 'phoneme'. You can override using the
## `.groups` argument.## Warning: position_dodge requires non-overlapping x intervals
Total Number of Tokens per phoneme and language after removal of outliers
Analysis of the included data
Figure @ref(fig:uplot) shows the productions of each of the four groups for the phonemes /u/ (in Spanish and English) and /y/ (in French and German). This condition was intended to explore how Spanish-English bilinguals would produce /y/, which is not present in either Spanish or English. An inspection of this figure suggests that /y/ was effectively imitated by all groups. In addition, it does not appear that there is much meaningful difference between the French and German imitations of the /y/ sound.
The models further supported the effective imitation of French and German /y/ by both bilingual groups. Figure @ref(fig:uplots) shows the conditional effects of the F1 and F2 models for /u/ in English and Spanish, and /y/ in French and German. Overall, the results of the models suggest that both bilingual groups performed similarly and could effectively imitate the French and German /y/. The models provided evidence that both bilingual groups had more fronted production of French and German /y/ relative to their Spanish and English /u/ (panel b). Additionally, although the English /u/ was slightly more fronted than Spanish /u/ by both groups, there was compelling evidence that productions English and Spanish /u/ German and French /y/ were distinct in terms of frontness, since there was no overlap in the posterior distributions.
Productions of /u/ in each language by each group
Conditional effects plot of /u/ for F1 (a) and F2 (b) in the bilingual groups
Figure @ref(fig:oplot) shows the productions of /o/ by all groups in Spanish, French and German, while the English sound was //. Although this condition was intended to bias the use of Spanish, this visualization does not imply that this was successful, nor that there were many meaningful differences between groups. It appears that all groups distinguish and similarly imitate the French and German /o/ sounds, regardless of whether they are bilingual or monolingual, or whether they differently produce the Spanish /o/ from the English //. Figure @ref(fig:oplots) again shows the conditional effects of the Bayesian models for F1 and F2. The models shows that both bilingual groups produced English // and French /o/ as more fronted than Spanish and German /o/ (panel b), while height varied by both groups as a function of language (panel a).
Productions of /o/ in each language by each group
Conditional effects plot of /o/ for F1 (a) and F2 (b) in the bilingual groups
Figure @ref(fig:iplot) shows the productions of /i/ in all languages and all groups. This condition was intended to examine how Spanish-English bilinguals produced novel language sounds that were present in both their L1 and L2. The figure does suggest that any of the groups produced /i/ distinctly as a function of language. The Bayesian analysis, however, revealed that the English L1 group had a more fronted /i/ in German and English than in Spanish and French (@ref(fig:iplots), panel a). The Spanish L1 group showed less variation in /i/ fronting as a function of language (@ref(fig:iplots), panel b). @ref(fig:iplots) also shows that both /i/ was slightly higher in German and Spanish by both groups.
Productions of /i/ in each language by each group
Conditional effects plot of /i/ for F1 (a) and F2 (b) in the bilingual groups
However, figure @ref(fig:aplot) shows the production of the wedge in English, French and German and /a/ in Spanish. This condition was intended to bias the production of English-like L3 vowels, with the Spanish /a/ for comparison. The results in this case are not consistent across groups. The L1 English bilingual group appeared to produce the German wedge, the English wedge, and the Spanish /a/ similarly. The L1 Spanish group produced the German and English wedge similarly, but diverged slightly in their production of /a/. All groups produced the French wedge similarly to one another, and dissimilarly from their own productions of other languages. @ref(fig:schwaplots) shows that both bilingual groups produced vowels similarly in a particular language when they produced the wedge or /a/ (in the case of Spanish).
Productions of /a/ and the wedge in each language by each group
Conditional effects plot of the wedge for F1 (a) and F2 (b) in the bilingual groups
VOT Production
Subset procedure
Participants
A total of 33 participants met subset criteria to be included in the VOT analysis (see appendix) Only 3 participants were L1 English speakers and 30 were L1 Spanish speakers.
VOT
In addition to vowels, the tokens which include stop consonants were analyzed for VOT. Each participant produced a total of 12 stops per language by producing three repititions of /pVf/ for the four distinct vowels.
Previous research in multilingualism has found that it is difficult to find a group of multilingual participants who produce sounds between languages distinctly enough to determine which language they could have come from, although they may have high proficiency. For example, when examining L3 French production by Spanish-Englsih bilingnuals, Parrish (2022) recruited XX Spanish-English bilinguals, but only XX produced English and Spanish distinctly and were included in the study. The inclusion of a given participant in Parrish (2022) was determined by segmenting all data from every participant and running a t-test with a participants Spanish and English. When this t-test returned a significant result, the participant was included.
Utilizing the context and insights of this previous research, the present dissertation subset participants in an effort to save resources in a replicable way. Rather than segmetning and subsquently discarding the acoustic data of discluded participants, the present dissertation used blinded judgment data by the author to determine which pariticpants were producing English and Spanish that were distinct enough to be informative when compared to their French and German productions. In particular, an R script was used to find all sound files that began with /p/ and were produced in each bilingual participant’s L2. From there, these files were copied to new directory and used to create a 2-alternative forced choice task in Psychopy. In total, 774 tokens were included, where 86 total participants (66 Spanish L1, 20 English L1) included 9 of their total 12 tokens each (language specific vowels /x/ and /y/ were discluded)Foot.
Procedure
One at a time, each sound was played, and “English-like” and “Spanish-like” were displayed on the left and right sides of the screen. After the sound was played, the researcher chose whether the segment /p/ sounded like an English-like or Spanish-like production of that sound, not knowing whether the speaker was an English or Spanish speaker. Between trials, a cross was displayed for 500ms in the center of the screen.
Results
Participants were included when more than half of their utterances were judged correctly. In total, 33 participants fit this criteria and had all of their data subsequently segmented (mean correct = 7.2 (SD = 1.5)). Overall, 3 of the participants were L1 English speakers and 30 were L1 Spanish speakers.
Footnote Although /x/ and /x/ are also language specific, the results of the first production experiment suggest that their production by these bilinguals is not distinct.
Analysis of VOT of the subset participants
The VOT of /p/ for each individual L1 English-L2 Spanish speaker
The VOT of /p/ for L1 Spanish-L2 English group
Adjusted credible intervals of all plausible values of VOT as a function of Language
@ref(fig:eng-vot) @ref(fig:span-vot) @ref(fig:vod-mod)
Discussion
- When bilinguals encounter a new language sound that is not phonemic in their L1 or L2, at least with the German and French /y/, they appear to have no problems imitating this sound, and do not assimilate the sound to their /u/ or /i/ categories in production.
No clear bilingual advantage was found in the analysis of the data.